JPS61249182A - Pattern recognizing and learning device - Google Patents

Abstract

PURPOSE:To execute efficiently the learning of a recognizing dictionary by controlling the number of an input pattern used for the learning of the recognizing dictionary of a special category, displaying a special category name, etc., to execute the learning and accelerating the input of a pattern. CONSTITUTION:A sound signal inputted through a voice input part 1 is converted to an input voice pattern at a characteristic extracting part 2. A recognizing part 3 executes a synthetic similar degree calculation between the characteristic vector of the input voice pattern and the dictionary pattern registered in a dictionary memory 4, and makes the category with the similar degree as a maximum into the recognizing result of the input voice pattern. On the other hand, at a memory part 6, the input voice pattern is successively accumulated and controlled. A learning part 8 learns the dictionary pattern for respective recognizing object categories by using the input voice pattern and the input voice pattern accumulated at the memory part 6. The learning part 8, in order to realize the effective learning, displays the category name and the number of the input pattern used for the learning up to the present through a displaying part 5, accelerated the input of the pattern necessary for the learning and executes the efficient learning.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field of the Invention] The present invention is directed to 1! used in input pattern recognition processing. The present invention relates to a pattern recognition learning device that can effectively learn idioms and improve its recognition performance.

[Technical background of the invention and its problems]

Conventional pattern recognition II devices perform some processing on several patterns input during learning and register them almost as-is as standard patterns in the ms dictionary memory. In many cases, a standard pattern and an input pattern are compared using, for example, the DP matching method, and the input pattern is recognized.

However, the standard pattern used in this type of DP matching method is! ! ! Since this method does not take into account the statistical distribution of speech targets, it is insufficient to be applied to, for example, speech recognition for unspecified speakers. Therefore, recently, i! The humble method is attracting attention.

The recognition dictionary used in this recognition method is created taking into consideration the statistical distribution of recognition patterns, but it is generally very difficult to create it. Therefore, the width of the above-mentioned statistical distribution is limited and many learning patterns are collected within that range to create, for example, an n-dimensional dictionary pattern. However, even with this method, creating the recognition dictionary (dictionary pattern) required a lot of effort and a lot of work time.

For these reasons, the reality is that those who have created recognition dictionaries use the same recognition dictionaries as they are for pattern recognition. In some cases, problems such as not being able to obtain a sufficient recognition rate occurred.

Therefore, in order to improve the recognition rate, l! Various methods for adapting recognition dictionaries have been proposed, but most of them simply modify the recognition dictionary to match a specific recognition target using a simple method. For this reason, we took into consideration the overall statistical distribution of recognition patterns! It was impossible to hope for an enriched recognition dictionary, and it was not possible to improve the recognition rate for recognition objects with the following distribution.

[Purpose of the invention]

The present invention has been made in consideration of these circumstances, and its purpose is to efficiently collect patterns necessary for learning a recognition dictionary used for pattern recognition, and to collect the patterns that are necessary for learning the recognition dictionary used for pattern recognition. An object of the present invention is to provide a pattern recognition learning device capable of effectively learning the internal dictionary and improving the WIII rate.

(Summary of the Invention) The present invention recognizes the input pattern by comparing the input pattern with a recognition dictionary using a composite similarity method or a subspace method, and uses the input pattern based on the i! recognition result. In a pattern recognition learning device equipped with a means for learning the W111 dictionary of a specific category, the number of input patterns used for learning the all dictionary of the specific category is managed, and the name of the specific category to be learned and the learning method are determined. The present invention is characterized in that the required number of pattern inputs and further data input examples including the specific category are displayed on the display to encourage input of the pattern of the specific category.

Then, for example, update the covariance matrix and KLI! l! of the specific category by changing the contribution rate of the input pattern for learning to the characteristic kernel or the covariance matrix using the open method. ! It is designed to perform learning processing for the Makoto dictionary.

[Effect 1 of the invention] Thus, according to the present invention, li! The operator is prompted to input the patterns necessary for learning aS dictionary by displaying the required number and input pattern examples, so the operator can easily and efficiently input the patterns necessary for learning aS dictionary 1 according to the displayed information. It becomes possible to input. Moreover, on the device side, the patterns necessary for learning the recognition dictionary can be efficiently collected, for example, by updating the covariance matrix and using the KL expansion method. ! ! It is possible to effectively perform learning that fully takes into account the distribution of m-knowledge objects in the knowledge dictionary.

Therefore,! ! ! This has great practical effects, such as effectively enriching the internal dictionary and improving its recognition rate.

[Embodiment of the Invention] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic configuration diagram of an example vRH. In addition, here is the input audio pattern X! I will explain about II, but pattern recognition such as handwritten character recognition! The same can be applied to knowledge.

The audio input unit 1 includes a microphone and an amplifier, and converts input audio into an electrical signal. The audio signal input through the audio input unit 1 is given to the feature extraction unit 2,
For example, the signal is frequency-resolved through an 8-channel filter bank and converted into an input audio pattern in which each frequency component is used as a characteristic parameter of the input audio.

! ! When the feature vector of the input speech pattern is f, the intelligence unit 3 calculates N')"EI"n < t, φl1ln)2
A composite similarity calculation of S"'[f]-Σ□n=1 11fl" is performed, and the category that maximizes the similarity S is set as the NII result of the input speech pattern. Note that this recognition processing method may of course be performed using Mahalanobis general distance, Euclidean distance, or the like.

The recognition result of the input voice pattern obtained in this way is displayed on the display section 5.

On the other hand, the input speech patterns obtained by the feature extraction section 2 are sequentially stored in the memory section.

This memory unit 6 stores a plurality of input speech patterns, and it is possible to check how many input speech patterns have been stored for each recognition target category, and which data of the accumulated input speech patterns have already been processed for learning. We manage each site to see if it has been used.

Furthermore, since the storage capacity of the memory unit 6 is limited, when the memory unit 6 becomes full, it is possible to control which data is discarded first or which data is stored in the external memory 7. There is.

The learning section 8 converts the dictionary patterns registered in the dictionary memory 4 into each I! using the input speech pattern obtained through the feature extraction section 2 and the input speech pattern stored in the memory section 6. The method is to learn for each category of objects to be recognized. Learning of dictionary patterns in this learning 81I8 is performed, for example, by repeatedly updating the covariance matrix and KL expansion of the covariance matrix data, as shown in FIG. That is, the learning is performed by performing the following calculation.

However, S is a learning pattern, and is given as a 64-dimensional vector, for example. Further, K is a covariance matrix before learning the pattern S, and K' is a covariance matrix after learning. W is a weighting coefficient that takes a positive or negative value; if it is positive, it acts to increase the similarity to the input pattern of each characteristic of the covariance matrix, and if it is negative, it acts to decrease the similarity. present.

Learning of such a covariance matrix is performed multiple times using multiple sample patterns for each recognition target category, and the covariance matrix vector obtained as a result of this learning is divided by KLjm and its eigenvalue is calculated. and the eigenvectors are calculated. The eigenvalues and their eigenvectors obtained through this learning and KLm expansion are registered in the dictionary memory 4 as a recognition dictionary for each recognition target category.

Furthermore, when the storage capacity of the dictionary memory 4 becomes full,
The recognition pattern for each category may be stored in the external memory 7 and loaded into the dictionary memory 4.

In this device, fill dictionaries (dictionary patterns) for each category are basically learned as described above. However, the KL expansion described above is executed using a calculation method such as the Jacobi method, the Householder-Sturm method, or the power law, but the calculation process generally requires a large amount of time. Therefore, the above-mentioned covariance matrix updating process is repeated as many times as necessary to obtain a certain degree of learning effect, and the covariance matrix vector is KL-expanded and registered in the dictionary memory 4. I try to do some learning.

In order to realize learning as effective as the recognition dictionary in this way, the learning unit 8 manages the number of input patterns required for the learning for each category, and also manages the number of input patterns that have already been used in the learning process. and manages the number of input patterns. Then, the learning category name and the number of input patterns used for learning the recognition dictionary up to now are displayed on the display unit 5 as shown in FIG. 3, and the patterns necessary for the learning are input. is encouraging. In this case, input voice pattern examples registered in the utterance example memory 9 are simultaneously displayed to display voice examples for obtaining input patterns that enable effective learning.

Specifically, when learning a recognition dictionary for speech category r, a, j, a minimum of 2011 input patterns are required for the learning, based on the weighting coefficient W mentioned above, the processing accuracy (bit precision) of the recognition device itself, etc. The learning unit 8 determines that this is the case. In addition, to determine the number of input patterns required for learning,
It is sufficient if it is determined according to the learning specifications of the recognition dictionary.
For example, it can be determined for each recognition target category using empirical knowledge about learning recognition dictionaries, or it can be determined that the recognition dictionary for each recognition target category is always trained using 10 or more input patterns. .

The learning section 8 displays the remaining number of input patterns required for learning from the number of input patterns already used for learning the recognition dictionary via the display section 5 according to the information on the number of necessary cover turns. ing. As a result,
The operator can know the number of input patterns required for learning the OII dictionary of the presented category.

At this time, if the learning category "AJ" is simply inputted as a single syllable, learning that takes into account the variation in the utterance pattern of "A" during continuous utterances will not be possible. In other words, a single syllable in speech exhibits relatively large fluctuations due to the influence of the syllables before and after it. Therefore, in order to absorb such fluctuations in syllable patterns, a word in which other syllables are added before and after the single syllable "a" is obtained from the utterance example memory 9, and this word is used as an input example of a speech pattern. It is displayed via the display section 5. For example, when learning an m-knowledge dictionary for the phonetic category raj, the user is prompted to vocalize words such as ``shiai'', ``kiatsu'', ``shiage'', etc. that include this ``a''. These words may be displayed selectively one by one, or a plurality of words may be displayed and sequentially specified using a cursor or the like. Alternatively, the operator may select and specify the specified word by himself from among the plurality of displayed words and input the specified word aloud.

However, when inputting speech for learning, the apparatus extracts the input pattern of the category to be learned from among the patterns of the input speech, stores it in the memory section 6, and provides it to the learning section 8. ing. As a result, the above-described learning of the recognition dictionary is executed according to the input pattern.

As explained above, according to the present device, the number of input patterns required for effective learning of the recognition dictionary is presented to the operator, and effective utterance examples (
Since the operator is instructed on the words (words), the operator only has to input the voice according to the instructions. Furthermore, since the learning section 8 effectively obtains only the input patterns necessary for learning the recognition dictionary and uses these to execute the learning process, it is possible to prevent the occurrence of learning errors due to undesired input patterns. do not have. Moreover, since only effective input patterns that can improve the learning effect can be efficiently collected in the required number, the recognition dictionary can be effectively trained.

Therefore, the recognition dictionary can be efficiently learned with less waste, and the recognition rate can be improved, which has great practical effects.

Note that the present invention is not limited to the embodiments described above. For example, the algorithm for determining the number of input patterns required for learning a recognition dictionary is i! It may be determined according to the structure of the dictionary. Another handwritten letter! l! ! It goes without saying that the present invention can be similarly applied to learning a recognition dictionary in a pattern recognition device such as a computer. Furthermore, the presentation method of pattern input necessary for learning can be modified in various ways. In short, the present invention can be implemented with various modifications without departing from the gist thereof.

[Brief explanation of the drawing]

Fig. 1 is a schematic diagram of a device according to an embodiment of the present invention, Fig. 2 is a diagram showing a learning algorithm of a recognition dictionary in the embodiment device, and Fig. 3 is an example of message output prompting pattern input in the embodiment device. FIG. 1...Audio input section, 2...Feature extraction section, 3...!
i! identification section, 4... dictionary memory, 5... display section, 6.
...Memory part, 1...External memory, 8...Learning part,
9... Utterance example memory. Applicant's agent Patent attorney Takehiko Suzue 113rl! J

Claims (4)

[Claims] (1) means for performing recognition processing on the input pattern by comparing the input pattern with a recognition dictionary; means for learning a recognition dictionary for a specific category using the input pattern based on the recognition result; 1. A pattern recognition learning device comprising means for managing the number of input patterns used for learning a recognition dictionary and prompting input of a pattern of the specific category. (2) The pattern according to claim 1, wherein the means for prompting input of a pattern of a specific category is one that displays the specific category name and the number of patterns required for learning the recognition dictionary of the specific category. Recognition learning device. (3) The pattern recognition learning device according to claim 1, wherein the means for prompting input of a pattern of a specific category is performed by displaying an example of data input including the specific category on a display. (4) The pattern recognition process by matching the input pattern with the recognition dictionary is performed by the composite similarity method or the subspace method, and the learning process of the recognition dictionary for a specific category involves updating the covariance matrix and KL 2. The pattern recognition learning device according to claim 1, wherein the pattern recognition learning device uses an expansion method and changes the contribution rate of the input pattern to the characteristic kernel or the covariance matrix.